DimensionSparsityMaskCreator creates unexpected mask_type about sparseml HOT 6 CLOSED

Hi @clementpoiret thank you for finding and fixing this bug. I've approved #442 and will get that landed.

We're currently overhauling our structure pruning support including for filter pruning which is reducing the wrong dimensions (should be [0] instead of [0,1]) and adding support for pruning groups of parameters together (#440).

from sparseml.

Thanks for the information @bfineran
As I am using sparseml to prune and quantize my models, does it mean I have to change (at least locally) [0,1] by [0] to have the expected filter pruning? Or is it better to wait until 0.9.0?

from sparseml.

Hi @clementpoiret, we'd recommend waiting until 0.9 for better-structured pruning support. This should go out in the next two weeks and we'll keep you updated as this progresses.

from sparseml.

Hi @clementpoiret
Release .9 has gone out the door; we hope you can try our latest version to address this issue; a summary of release highlights is also available - check them out here.

Best, Jeannie / Neural Magic

from sparseml.

Hi @clementpoiret to follow up on the comment from @jeanniefinks, we have added a StructuredPruningModifier in the 0.9 release for PyTorch. Let us know if you have a chance to try it out as this enables proper filter pruning. Happy to help walkthrough an example as well!

Thanks,
Mark

from sparseml.

I launched another training using StructuredPruningModifier to prune filters, but some questions are coming to my mind:
1/ In the doc, GMPruningModifier still have "filters" and "channels". Is it now equivalent to using the StructuredPruningModifier with the same arguments?
2/ How can I check if the model is correctly pruned? The final pruned ONNX file weights the same thing (around 140Mb) than its non-pruned counterpart, is this normal?

Thanks!

from sparseml.